Facial expressivity in schizophrenia has been a topic of clinical interest for the past century. Besides the schizophrenia sufferers difficulty decoding the facial expressions of others, they often have difficulty encoding facial expressions. Traditionally, evaluations of facial expressions have been conducted by trained human observers using the facial action coding system. The process was slow and subject to intra and inter-observer variability. In the past decade the traditional facial action coding system developed by Ekman has been adapted for use in affective computing. Here we assess the applications of this adaptation for schizophrenia, the findings of current groups, and the future role of this technology.

We review the applications of computer vision technology in schizophrenia using pubmed and google scholar search criteria of “computer vision” AND “Schizophrenia” from January of 2010 to June of 2018.

Five articles were selected for inclusion representing 1 case series and 4 case-control analysis. Authors assessed variations in facial action unit presence, intensity, various measures of length of activation, action unit clustering, congruence, and appropriateness. Findings point to variations in each of these areas, except action unit appropriateness, between control and schizophrenia patients. Computer vision techniques were also demonstrated to have high accuracy in classifying schizophrenia from control patients, reaching an AUC just under 0.9 in one study, and to predict psychometric scores, reaching pearson’s correlation values of under 0.7.

Our review of the literature demonstrates agreement in findings of traditional and contemporary assessment techniques of facial expressivity in schizophrenia. Our findings also demonstrate that current computer vision techniques have achieved capacity to differentiate schizophrenia from control populations and to predict psychometric scores. Nevertheless, the predictive accuracy of these technologies leaves room for growth. On analysis our group found two modifiable areas that may contribute to improving algorithm accuracy: assessment protocol and feature inclusion. Based on our review we recommend assessment of facial expressivity during a period of silence in addition to an assessment during a clinically structured interview utilizing emotionally evocative questions. Furthermore, where underfit is a problem we recommend progressive inclusion of features including action unit activation, intensity, action unit rate of onset and offset, clustering (including richness, distribution, and typicality), and congruence. Inclusion of each of these features may improve algorithm predictive accuracy.

We review current applications of computer vision in the assessment of facial expressions in schizophrenia. We present the results of current innovative works in the field and discuss areas for continued development.